The Wedge Is Not the Mission

In May I wrote a piece called The PM With Loaded Dice. A friend had asked whether I would want to be a product manager at OpenAI or Anthropic . I told him I would want the job more than almost any other PM seat. [1]

I remember being a product manager at Applied Biosystems. We surveyed customers, watched competitors, ran the numbers and tried to read the market. Mostly, we guessed.

The product managers at the labs do not have to guess the way we did. OpenAI and Anthropic publish privacy-protected slices of how people use their models through OpenAI Signals and the Anthropic Economic Index. [2][3] The internal view is almost certainly richer, including which features drive usage and revenue.

That was the loaded dice.

The dashboard does not just tell the product manager where demand is going. Once enough money is attached, it can start telling the company where to go.

For most companies, that is exactly what should happen. Listen to customers. Solve their problems. Follow the money.

The complication is what these labs say they are building.

OpenAI says its mission is to ensure that AGI benefits all of humanity. [4] Anthropic says it is trying to build frontier AI systems that are reliable, interpretable and steerable. [5] Google DeepMind says its mission is to build AI responsibly to benefit humanity. [6]

The wording is different. The promise is bigger than just better software. Much bigger.

What happens when the next thing a customer wants is not the next thing that moves the mission forward?

The Wedge Matters

The work inside spreadsheets, documents, coding tools and email matters.

I have spent the past few years writing The Invisible Interface. The book is about what happens when AI stops sitting in a chat window and starts working inside the systems people already use. It needs context, memory, permissions, audit trails and the ability to act. That is when it becomes an operating layer.

So I am not dismissing the enterprise work. I think it is how the technology gets adopted.

Consumer products matter too. They put the models in front of more people and more tasks. They show the labs where the systems are useful and where they fail.

OpenAI launched ChatGPT Enterprise with enterprise privacy and security, deployment tools, advanced data analysis and customization. [7] Anthropic markets Claude for coding, research, document creation and company workflows. [8] Microsoft’s Copilot can now take multistep, app-native actions inside Word, Excel and PowerPoint. [9]

That is the wedge. It gets AI into daily work, generates revenue and feedback, and gives the system a chance to earn trust.

But customers do more than fund the roadmap. What they use, complain about and pay for can shape what gets improved.

If you have ever built for enterprise customers, you know the pull. A bank needs more controls. A pharmaceutical company needs secure access to its data. A finance team needs better spreadsheet reasoning. Nothing is wrong with any of that. But once enough revenue sits behind a request, it is no longer just a request. It is a commitment.

The mission statement can stay exactly the same while the company underneath it changes.

To be clear, I do not know that this is happening inside the frontier labs. The public record does not show how they divide their best people, compute and leadership attention.

And the labs have not walked away from science.

The AlphaFold Protein Structure Database, developed by Google DeepMind and EMBL-EBI, provides more than 200 million protein structure predictions. [10] OpenAI introduced GPT-Rosalind as a research preview for qualified U.S. Enterprise customers working in the life sciences. [11] Anthropic released Claude Science in beta, with auditable histories behind its outputs intended to make the work easier to validate and reproduce. [12]

So science is still there. But those examples do not tell us what happens when improving the underlying model and shipping the next commercial product require the same people, compute and money.

Which one gets protected?

Does the Wedge Strengthen the Machine?

The easy argument is that every engineer improving a business product is one fewer working on the underlying model. The economics are not that simple.

In a dashboard updated February 5, 2026, Epoch AI reported that frontier labs had raised more than $170 billion. It also estimated that training compute for frontier language models had grown fivefold per year since 2020, while the cost of training them had grown 3.5-fold per year. The capital problem is glaringly obvious. [13]

Commercial products can pay for compute and keep the research going. Some of the same work – coding, evaluation and tool use – can improve the models too.

The question is not product or research. It is which product work makes the underlying system better, and which work simply makes the business larger.

Now add one more thing: AI is starting to help build the next generation of AI.

Recursive self-improvement is the point where an AI system can design and build a more capable successor on its own. We are not there yet. Anthropic says it is delegating a growing share of AI development to AI systems. For now, humans still largely choose the problems and decide which results matter. [14]

OpenAI’s GPT-5.6 system card, published July 9, 2026, says all three models in the family remain below its Preparedness Framework’s “High” capability threshold for AI Self-Improvement. [15]

The version happening now looks less like science fiction. Models can write and test code, run experiments and help search for more efficient algorithms. They are taking over pieces of the work required to build the next AI.

Google DeepMind’s AlphaEvolve is the cleanest example. It combines Gemini models with automated evaluators that score candidate programs, then uses an evolutionary process to improve promising ideas. DeepMind reports that one AlphaEvolve change sped up a matrix multiplication kernel in Gemini’s architecture by 23%, leading to a 1% reduction in Gemini’s training time. It also reports that a scheduling heuristic discovered by AlphaEvolve continuously recovers, on average, 0.7% of Google’s worldwide compute resources. [16]

That is AI-assisted AI development. It is not a system autonomously designing its own successor. But at Google’s scale, a 1% reduction in training time is real money. Later training runs that continue to use the kernel can inherit the improvement.

A better spreadsheet feature improves one workflow. A better way to build models can improve the models that follow.

Could Focus Win?

Could a smaller team focused on nothing but automating AI research get to a real self-improvement loop first?

Maybe. We do not know.

The large labs have enormous advantages: compute, specialized talent, mature infrastructure and access to their own frontier models. AlphaEvolve shows that scale and focus can live inside the same company. A smaller group would have to make focus count.

We can see what focus buys in science. Periodic Labs is building AI scientists and autonomous laboratories in the physical sciences. [17] FutureHouse is building AI agents to automate research in biology and other complex sciences. [18]

Neither is an AGI competitor. That is the point. Each is built around a narrower definition of progress and fewer competing demands.

Hypothesis. Experiment. Result. Repeat.

The company that owns the assistant may not be the company that builds the fastest scientific learning loop. A frontier lab could win both. I would not assume that.

What I Would Ask the Board

Apollo had a finish line. Kennedy told the country what the goal was: land a man on the Moon, return him safely to Earth and do it before the decade was out. The goal was unambiguous. [19]

The AI labs do not have a comparable deadline or a single measurable test for completion. Their public missions leave management a great deal of room to describe a successful product as progress toward the mission. That flexibility is useful. It also makes drift harder to see.

If I were sitting on one of these boards, I would want answers to three questions.

1: What does progress toward the mission mean when you strip out revenue, usage and customer growth?

2: How much of the most advanced compute and senior talent is protected for making the models smarter and safer, including work that may not have a buyer for years?

3: When a major customer request conflicts with the long-term research agenda, who has the authority to say no?

The exact numbers can stay private. The board cannot afford to be unclear about the answers.

The enterprise and consumer products may be exactly what pays for the machine and makes it reliable enough to matter in science, medicine, energy and materials. I am not arguing against them.

If I were the product manager, I would follow the signal. That is the job.

The board has a different job: make sure the signal does not become the mission.

References

[1] Harry Glorikian, “The PM With Loaded Dice,” May 17, 2026. https://glorikian.com/the-pm-with-loaded-dice/

[2] OpenAI, “OpenAI Signals.” https://openai.com/signals/

[3] Anthropic, “Anthropic Economic Index,” updated June 26, 2026. https://www.anthropic.com/economic-index

[4] OpenAI, “About.” https://openai.com/about/

[5] Anthropic, “Company.” https://www.anthropic.com/company

[6] Google DeepMind, “About Google DeepMind.” https://deepmind.google/about/

[7] OpenAI, “Introducing ChatGPT Enterprise,” August 28, 2023. https://openai.com/index/introducing-chatgpt-enterprise/

[8] Anthropic, “Claude Enterprise Plan.” https://claude.com/solutions/enterprise

[9] Microsoft, “Copilot’s agentic capabilities in Word, Excel, and PowerPoint are generally available,” April 22, 2026. https://www.microsoft.com/en-us/microsoft-365/blog/2026/04/22/copilots-agentic-capabilities-in-word-excel-and-powerpoint-are-generally-available/

[10] Google DeepMind and EMBL-EBI, “AlphaFold Protein Structure Database.” https://alphafold.ebi.ac.uk/

[11] OpenAI, “Introducing GPT-Rosalind for life sciences research,” April 16, 2026. https://openai.com/index/introducing-gpt-rosalind/

[12] Anthropic, “Claude Science, an AI workbench for scientists, is now available,” June 30, 2026. https://www.anthropic.com/news/claude-science-ai-workbench

[13] Epoch AI, “Trends in Artificial Intelligence,” updated February 5, 2026. https://epoch.ai/trends

[14] Anthropic, “When AI builds itself.” https://www.anthropic.com/institute/recursive-self-improvement

[15] OpenAI, “GPT-5.6 System Card,” section “AI Self-Improvement Capabilities,” July 9, 2026. https://deploymentsafety.openai.com/gpt-5-6/ai-self-improvement-capabilities

[16] Google DeepMind, “AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms,” May 14, 2025. https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/

[17] Periodic Labs, “Accelerate science.” https://periodic.com/

[18] FutureHouse, “Automating scientific discovery.” https://www.futurehouse.org/

[19] NASA History Office, “The Decision to Go to the Moon: President John F. Kennedy’s May 25, 1961 Speech before a Joint Session of Congress,” September 22, 1998. https://www.nasa.gov/history/the-decision-to-go-to-the-moon/